我正在尝试使用C ++ API在Tensorflow中运行从Keras导出的模型。该模型已正确导出和导入,但是当我尝试在C ++中运行会话时,我在Run
的{{1}}成员中传递的变量名称被修改,从而导致错误:
tensorflow::Session
这是我在Keras中建立的模型:
Invalid argument: Tensor Output0:0, specified in either feed_devices or fetch_devices was not found in the Graph
使用S = Input(shape=[state_size], name='Input')
h0 = Dense(self.HIDDEN1_UNITS, name='Hidden1', kernel_initializer=keras.initializers.RandomUniform())(S)
h0_a = Activation('relu')(h0)
h1 = Dense(self.HIDDEN2_UNITS, name='Hidden2', kernel_initializer=keras.initializers.RandomUniform())(h0_a)
h1_a = Activation('relu')(h1)
output = Dense(1)(h1_a)
output_a = Activation('tanh', name='Output0')(output)
model = Model(S,output_a)
导出模型
和tensorflow.python.framework.graph_util
,遵循https://github.com/amir-abdi/keras_to_tensorflow
然后使用C ++ API将模型导入Tensorflow:
tensorflow.python.framework.graph_io
该图似乎已成功导入,实际上,当我在Tensorflow中打印图中包含的节点名称时,会得到以下信息:
ParseProtoUnlimited(&graph_def, <graph_protoprotobuffer>, <graph_size> );
现在,如果我想尝试运行该会话,请执行以下操作:
Names : Input_5
Names : Hidden1_5/kernel
Names : Hidden1_5/kernel/read
Names : Hidden1_5/bias
Names : Hidden1_5/bias/read
Names : Hidden1_5/MatMul
Names : Hidden1_5/BiasAdd
Names : activation_5_1/Relu
Names : Hidden2_5/kernel
Names : Hidden2_5/kernel/read
Names : Hidden2_5/bias
Names : Hidden2_5/bias/read
Names : Hidden2_5/MatMul
Names : Hidden2_5/BiasAdd
Names : activation_6_1/Relu
Names : dense_3_1/kernel
Names : dense_3_1/kernel/read
Names : dense_3_1/bias
Names : dense_3_1/bias/read
Names : dense_3_1/MatMul
Names : dense_3_1/BiasAdd
Names : Output_5/Tanh
Names : output0
但是应用程序抛出以下错误:
Tensor input(DT_FLOAT, TensorShape( { 4 }));
...
[init tensor]
...
session->Run({{"Input_5", input}}, {"Output0"}, {},
&outputs);
为什么函数Run session将:0附加到要检索的输出张量上?
此外,我已经尝试过使用Tensorflow构建的简单图形进行上述操作,并且C ++上的会话运行没有问题。这是起作用的图:
Invalid argument: Tensor Output0:0, specified in either feed_devices or fetch_devices was not found in the Graph
with tf.Session() as sess:
a = tf.Variable(5.0, name='a')
b = tf.Variable(6.0, name='b')
c = tf.multiply(a, b, name="c")
sess.run(tf.global_variables_initializer())
exported = sess.graph_def.SerializeToString()
size = sess.graph_def.ByteSize()
为什么在第二种情况下我设法运行该会话,而在第一种情况下却遇到该错误?
Keras版本为Tensor a(DT_FLOAT, TensorShape());
a.scalar<float>()() = 5.0;
Tensor b(DT_FLOAT, TensorShape());
b.scalar<float>()() = 6.0;
std::vector<std::pair<string, tensorflow::Tensor>> inputs = {
{ "a", a }, { "b", b }, };
std::vector<tensorflow::Tensor> outputs;
Run the session, evaluating our "c" operation from the graph
status = session->Run(inputs, { "c" }, { }, &outputs);
Tensorflow版本为2.2.2
任何帮助将不胜感激。